Many GB5 (and real world) tasks are memory bandwidth bottlenecked, which greatly favors M1 Max because it has over double a Threadripper's memory bandwidth.

Sort of. The CPU complex of the M1 Max can achieve ~200 GB/s, you can only hit the 400 GB/s mark by getting the GPU involved.

At the same time the Threadrippers also have a gargantuan amount of cache that can be accessed at several hundred gigabytes per second per core. Obviously not as nice as being able to hit DRAM at that speed.

That cache is not uniform time access. It costs over 100ns to cross the IO die to access another die's L3, almost as much as going to main memory. In practice you have to treat it as 8 separate 32 MB L3 caches.

Also, not everything fits into cache.

Yeah, it’s the real world tasks that GeekBench tries to simulate that don’t tend to scale linearity with processor count. A lot of software does not take good advantage of multiple cores.

That's certainly true. But if that's your workload you shouldn't be buying a 64-core CPU...

I use a few 32 and 64 core machines for build servers and file servers, and while the 64-core EPYCs are not twice as fast as the 32-core ones due to lower overall frequency, they're 70% or so faster in most of the things I throw at them.

Does Geekbench actually attempt to simulate that in their multi-core score? And how?

I was under the impression that all of their multi-core tests were "run N independent copies of the single-threaded test", just like SPECrate does.

> A lot of software does not take good advantage of multiple cores.

It sounds pointless to come up with synthetic benchmarks which emulate software that is not able to handle hardware, and then use said synthetic benchmarks to evaluate the hardware performance.

It has a very specific point: communicating performance to people who don't know hardware.

Most consumers are software aware, not hardware aware. They care what they will use the hardware for, not what they can use it for. To that end, benchmarks that correlate with their experience are more useful than a tuned BLAS implementation.

Probably it's the thermals that don't scale. The more the cores, the lower the the peak performance per core.